psann 0.9.17

PSANN: Parameterized Sine-Activated Neural Networks (sklearn-style, PyTorch backend)

PSANN - Parameterized Sine-Activated Neural Networks

Sklearn-style estimators built on PyTorch that use learnable sine activations, optional persistent state, and shared helpers for episodic (HISSO) training.

Quick links:

• API reference: docs/API.md
• Technical design notes: TECHNICAL_DETAILS.md
• Scenario walkthroughs: docs/examples/README.md

Installation

python -m venv .venv
.\.venv\Scripts\Activate.ps1   # Windows PowerShell
# source .venv/bin/activate     # macOS/Linux
pip install --upgrade pip
pip install -e .                # editable install from source

Extras defined in pyproject.toml:

• psann[sklearn]: adds scikit-learn for real BaseEstimator mixins and metrics
• psann[viz]: plotting helpers used in notebooks/examples
• psann[dev]: pytest, ruff, black

Need pre-pinned builds (e.g. on Windows or air-gapped envs)? Use the compatibility constraints:

pip install -e . -c requirements-compat.txt

pyproject.toml is the authoritative dependency list. requirements-compat.txt mirrors the newest widely available wheels for NumPy, SciPy, and scikit-learn when you need lockstep installs.

NumPy 2.0 status

• Core PSANN modules run cleanly on NumPy 2.0; no removed aliases (e.g. np.bool, np.int) are referenced.
• The base install now depends on numpy>=1.23 with no upper cap. Use requirements-compat.txt when you need the older NumPy 1.26 wheels for legacy environments.
• The optional psann[sklearn] extra still uses scikit-learn<1.5; those wheels currently target NumPy 1.x. Install scikit-learn>=1.5 when available or stick to the compatibility constraints if you rely on that extra.

Quick Start

Supervised regression

import numpy as np
from psann import PSANNRegressor

rs = np.random.RandomState(42)
X = np.linspace(-4, 4, 1000, dtype=np.float32).reshape(-1, 1)
y = 0.8 * np.exp(-0.25 * np.abs(X)) * np.sin(3.5 * X)

model = PSANNRegressor(
    hidden_layers=2,
    hidden_units=64,
    epochs=200,
    lr=1e-3,
    early_stopping=True,
    patience=20,
    random_state=42,
)
model.fit(X, y, verbose=1)
print("R^2:", model.score(X, y))

Supervising extras outputs

Append extra columns to y or pass them separately. Extra heads are scheduled automatically when extras>0.

extras = np.stack([np.cos(X[:, 0]), np.sin(X[:, 0])], axis=1).astype(np.float32)
y_with_extras = np.concatenate([y[:, None], extras], axis=1)

model = PSANNRegressor(hidden_layers=2, hidden_units=64, extras=2)
model.fit(X, y_with_extras, verbose=1)

For streaming/time-series, LSM preprocessors, segmentation heads, and HISSO workflows, head to docs/examples/README.md.

Feature Highlights

• Learnable sine activations (SineParam) with amplitude, frequency, and decay bounds
• Shared _fit helper powering PSANN, residual PSANN, and language-model estimators
• Optional predictive extras with automatic target detection and rollout utilities
• Stateful controllers for streaming inference with warm-start and reset policies
• Convolutional variants that preserve spatial structure and support per-element outputs
• HISSO episodic training with reward hooks, supervised warm starts, and extras scheduling

Contributing

• Run pip install -e .[dev] and then pytest / ruff check .
• Format code with black and keep docstrings ASCII unless a file already uses Unicode
• See open TODO items in TODO.txt